GPT-Live-1, the full-duplex voice model OpenAI put into ChatGPT in July, became available to developers on 10 September 2026 at $0.05 a minute. On the same day OpenAI opened the Agents API as a public beta, offering the Codex harness as a hosted service. The two launches were written as a pair. GPT-Live-1 is the voice that talks to a customer, and the Agents API is one of the backends it can hand the real work to.
Announcements about AI models and AI agents are written to be read in a minute. The developer documentation takes longer, and it is where the limits live. We read both launch posts against 21 pages of OpenAI’s own documentation, including the model page, the price list, the data-controls table and the migration guide. We also pulled the raw data behind the seven benchmark charts in the GPT-Live-1 announcement.
Three findings stand out. Four of those seven charts score GPT-Live-1 with a GPT-6 Astra or Terra backend attached, against older voice models running alone. The $0.05 rate bills every open second, silence included. And the pair does not share data controls: GPT-Live-1 sessions qualify for Zero Data Retention and European processing, while the Agents API qualifies for neither. Our earlier pieces covered GPT-Live’s ChatGPT debut and the Managed Agents leak before DevDay. This one is about what developers actually get.
Table of contents
- GPT-Live-1 in the API: A Voice Model That Listens While It Talks
- GPT-Live-1 Benchmarks: Four of Seven Charts Carry an Asterisk
- What GPT-Live-1 Costs: The Meter Runs During Silence
- The Agents API: The Codex Harness as a Hosted Service
- Agents API Costs, Limits and Defaults Worth Knowing
- GPT-Live-1 and the Agents API Do Not Share Data Controls
- What GPT-Live-1 and Agents API Customers Claim
- Migrating a Realtime Voice Agent to GPT-Live-1
- Where This Leaves DevDay and Agent Builder
- Should Your Business Build on GPT-Live-1 and the Agents API?
- GPT-Live-1 and Agents API: Frequently Asked Questions
- References and Further Reading
GPT-Live-1 in the API: A Voice Model That Listens While It Talks
A voice agent has always had to wait its turn. GPT-Live-1 is OpenAI’s attempt to remove that wait, and the API release is the first time businesses can build on it directly.
Two launches, one day
OpenAI published “Build more natural voice experiences with GPT-Live-1 in the API” and “Introducing the Agents API” on 10 September 2026. The GPT-Live-1 release came 64 days after GPT-Live first shipped in ChatGPT on 8 July, when OpenAI said it planned to bring the models “to the API soon”. Yelp announced a production deployment the same day, with a supporting quote from Teri Yu, OpenAI’s multimodal product lead.
Full duplex, not turn-taking
Older voice agents take turns. A cascaded system chains speech-to-text, a language model and text-to-speech, and a turn-based speech model waits for silence before it answers. OpenAI’s July post noted that silence-based detection means “even a brief pause or background noise could be mistaken for the end of turn”. GPT-Live-1 processes incoming audio while it generates outgoing audio, so it can say “mhmm”, stay quiet while someone thinks, or stop when interrupted.
Delegation: the voice layer hands work to a backend
GPT-Live-1 is deliberately not the brain of the application. OpenAI’s getting-started guide splits the job in two: “GPT-Live handles conversation” while “the backend handles delegated tasks”. With Responses delegation, GPT-Live-1 calls a hosted Responses model that you configure. With client delegation, your own application runs any model, agent harness or service and passes the result back. The launch post suggests GPT-5.6 Luna for high-volume jobs such as order updates and GPT-6 Astra for complex customer issues.
What an interruption does not cancel
One sentence in the guide matters for anyone taking bookings or payments by voice: “Interrupting speech does not automatically cancel backend work.” If a caller says “actually, make it Friday” while a Thursday booking is in flight, that booking may still complete. The migration guide tells developers to track task revisions, invalidate earlier confirmations and reconcile any action that already succeeded. Your application, not GPT-Live-1, “owns permissions, confirmations, private function execution, and durable task state”.
GPT-Live-1 against GPT-Realtime-2.1, spec for spec
The model pages show GPT-Live-1 is a separate product line rather than a Realtime upgrade. It has its own endpoint, bills by time rather than tokens, and is rate-limited by concurrent sessions.
| Specification | GPT-Live-1 | GPT-Realtime-2.1 |
|---|---|---|
| API endpoint | v1/live/sessions only | v1/realtime only |
| Input and output | Audio and text in, audio and text out | Text, audio and image in, text and audio out |
| Knowledge cutoff | 31 July 2025 | 30 September 2024 |
| Context window | 128,000 tokens | 128,000 tokens |
| Reasoning | Delegated to a backend model | Built in, configurable effort |
| Billing | $0.05 per session minute, billed per second | Per token: audio $32 in and $64 out per million |
| Rate limits | 25 to 500 concurrent sessions by tier | Requests and tokens per minute |
| Function calling | Supported | Supported |
| Structured outputs | Not supported | Not listed |
The newer knowledge cutoff matters less than it looks, because GPT-Live-1 is meant to fetch facts from its backend. The absence of structured outputs matters more. Anything that has to arrive as clean, typed data belongs in the backend call, not the voice model. GPT-Live-1 also ships with 12 new voices, from Quartz to Cinder.
GPT-Live-1 Benchmarks: Four of Seven Charts Carry an Asterisk
OpenAI’s launch post makes two headline claims. GPT-Live-1 “improves Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1”, and paired with GPT-6 Astra at medium reasoning effort it “ranks #1 on Tau3”. The charts behind those claims are drawn from data embedded in the page, so we extracted all seven and checked the footnotes.
Every GPT-Live-1 chart, with its backend
The last column is the one most coverage left out.
| Benchmark | GPT-Live-1 | GPT-Realtime-2.1 | GPT-Realtime-2 | Backend behind the GPT-Live-1 score |
|---|---|---|---|---|
| Tau3 voice intelligence | 86.2% | 45.7% | 42.4% | Astra, medium effort |
| Tau Banking voice knowledge | 32.0% | 12.4% | 10.3% | Astra, medium effort |
| Full Duplex Bench v3 tool calling | 87.0% | 60.0% | 58.0% | Terra, low effort |
| Full Duplex Bench v3 response quality | 90.0% | 88.0% | 81.0% | Terra, low effort |
| Full Duplex Bench v1.5 interactivity | 80.1% | 45.4% | 47.8% | None |
| Full Duplex Bench v1 turn-taking latency | 0.798 s | 1.41 s | 1.63 s | None |
| Artificial Analysis conversational dynamics | 97.3% | 95.7% | 95.3% | None |
What the asterisk means
Four bars are labelled “gpt-live-1*”, with a footnote naming the backend: Astra at medium effort for the two Tau tests, Terra at low effort for the two Full Duplex Bench v3 tests. The GPT-Realtime bars carry no asterisk. On those four charts OpenAI is comparing a voice model plus a separate reasoning model against a single voice model working alone. That is an honest picture of how GPT-Live-1 is meant to be deployed, and the footnotes disclose it. It is not a like-for-like test of the two voice models. The Decoder, for one, reported the 87% tool-calling and 32% banking scores without mentioning the backend.
The GPT-Live-1 gains, ranked
The biggest gains are the ones measured with a backend attached; the smallest is a voice-only measure where the older models already scored above 95%.
The voice-only GPT-Live-1 results
Three charts test GPT-Live-1 with no backend, and these are the cleanest comparison. On interactivity it scored 80.1% against 45.4%. It started replying in 0.798 seconds against 1.41 seconds, about 43% sooner. On Artificial Analysis conversational dynamics the gap was 1.6 points. One detail cuts the other way for the older line: GPT-Realtime-2.1 scored lower than GPT-Realtime-2 on interactivity, 45.4% against 47.8%.
The 30-point claim checks out
On the Full Duplex Bench v1.5 chart, 80.1 minus 45.4 is 34.7 points. OpenAI’s “30 percentage points” understates the interactivity gap rather than inflating it. The claim that deserves the asterisk is the Tau3 ranking, because it is a result for GPT-Live-1 and GPT-6 Astra together.
GPT-Live-1 passed 31 of 97 banking tasks
OpenAI defines the Tau Banking score as “the fraction of 97 banking_knowledge tasks completed successfully”. A 32.0% pass rate is 31 tasks. GPT-Realtime-2.1 managed 12 and GPT-Realtime-2 managed 10. The best configuration OpenAI tested, GPT-Live-1 with Astra, still failed 66 of 97. For a bank that is an argument for a well-designed human handoff, not against voice. The system card also carries a 4 August correction: its safety evaluations had used “a backend configuration that did not match the final model release” and were re-run.
What GPT-Live-1 Costs: The Meter Runs During Silence
The launch post gives one price, $0.05 per minute “for the front-end voice layer”. The pricing page and the cost guide explain what that minute includes, and it is not only talking.
$0.05 a minute, billed by the second
GPT-Live-1 sessions “are billed per second, without rounding up to a whole minute. Backend model and tool usage is charged separately.” That makes an hour of open session $3.00 before any reasoning model is called. There is no token counting on the voice side, which makes GPT-Live-1 far easier to forecast than a token-billed speech model. The trade is that you pay for time rather than speech.
What counts as billable GPT-Live-1 time
OpenAI’s cost guide is explicit: active session time “includes time when the user speaks, the assistant speaks, both are silent, or the backend is working”. Muting the microphone does not close the session. Creating a WebRTC session bills 15 seconds while it initialises, credited once it starts, so reconnect loops cost money. The guide suggests closing the voice session during long backend tasks, noting that “closing saves $0.05 per minute of idle voice time”.
GPT-Live-1 against Realtime token rates
Realtime bills audio as tokens: one token per 100 milliseconds of user speech and one per 50 milliseconds of assistant speech, which is 600 and 1,200 tokens a minute. At OpenAI’s list rates, a minute of GPT-Realtime-2.1 speaking costs more than a GPT-Live-1 minute, but a minute of silence costs Realtime nothing once voice activity detection filters it out.
A four-minute booking call, priced three ways
Take a restaurant booking lasting 240 seconds: the caller speaks for 90 seconds, the agent for 75, and 75 seconds pass in pauses and lookups. These are voice charges only, calculated from OpenAI’s published rates.
| Line item | GPT-Live-1 | GPT-Realtime-2.1 | GPT-Realtime-2.1-mini |
|---|---|---|---|
| 90 seconds of caller speech | In session time | 900 tokens: $0.0288 | 900 tokens: $0.0090 |
| 75 seconds of agent speech | In session time | 1,500 tokens: $0.0960 | 1,500 tokens: $0.0300 |
| 75 seconds of silence and waiting | In session time | $0 | $0 |
| Voice charge | 240 seconds: $0.2000 | $0.1248 | $0.0390 |
| Billed on top | Backend model and tool tokens | Conversation history re-sent as input each turn, plus text tokens | Same as GPT-Realtime-2.1 |
The Realtime figures are floors, not quotes: every Realtime response re-sends the whole conversation as input, and later turns cost more. Even so, the arithmetic says GPT-Live-1 is not automatically the cheaper voice layer. It is the more predictable one, and it charges for silence that Realtime ignores. Call centres with long holds will feel that difference first.
GPT-Live-1 concurrency caps, and what full load costs
GPT-Live-1 is rate-limited by concurrent sessions, and the Free tier is not supported. At $3.00 per session hour, the cap also sets the maximum hourly voice bill.
A new account at Tier 1 can hold 25 calls at once. That is enough for a pilot on one phone line, not for a seasonal peak, so check your tier before launch day.
The Agents API: The Codex Harness as a Hosted Service
The second launch is aimed at a different problem. Long-running agents need a loop that manages context, calls tools and coordinates helpers, plus somewhere safe to run code. OpenAI now rents out the loop it built for Codex.
One call creates an agent
A single POST /v1/agents/sessions request with the header OpenAI-Beta: agents=v1 defines the model, instructions, tools and environment, and starts work. API keys need the api.agents.read, api.agents.write and api.responses.write permissions. The documentation describes four building blocks: the agent, an optional environment, a durable session, and the events and items that flow through it. Sessions can be steered mid-turn and continued later.
Three places the work can run
Every session picks none, openai_hosted or self_hosted. A hosted sandbox is a Linux workspace with Python, Node.js and command-line tools, and files saved under /workspace/outputs become downloadable artifacts. For self-hosted, you run codex exec-server inside your own compute, which connects outbound to OpenAI with a restricted key. The architecture page says an environment “can be a remote sandbox, your laptop, a Docker container, or an AWS Lambda function”.
What the managed harness does
OpenAI lists automatic compaction of long context, tool search that loads tool definitions only when needed, programmatic tool calling, MCP servers, web search, subagents and the ability to resume a session. The launch post promises “versioned access to these capabilities with each model launch”. The harness itself is the open-source Codex project, published under the Apache-2.0 licence and carrying 123,313 GitHub stars on 11 September.
Nine sandbox partners, up from seven in April
When OpenAI updated the Agents SDK on 15 April, it named seven built-in sandbox providers: Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop and Vercel. The Agents API adds DigitalOcean and Oracle. The same April post warned that “managed agent APIs can simplify deployment but constrain where agents run and how they access sensitive data”. Some 148 days later OpenAI shipped one, with self-hosted sandboxes as its answer to its own objection. We explored that trade-off in where your agents actually run.
Agents API, Agents SDK or Responses API
OpenAI’s own comparison rates integration effort as low for the Agents API, medium for the Agents SDK and high for the raw Responses API. The difference is who runs the loop. With the Agents API, OpenAI runs the harness and stores session state. With the SDK, the loop runs inside your application and your storage. Teams that need full control of state and logging will still find the SDK the better fit.
Agents API Costs, Limits and Defaults Worth Knowing
The launch post’s pricing line is short, and it is accurate as far as it goes. The overview page adds the parts a finance team will ask about.
“No additional fees” is not the same as tokens only
The announcement says there are “no additional fees for using the Agents API – you simply pay for the tokens and tools your agents use”. The overview spells out three meters: model usage at the model’s API rates, built-in tools at their standard rates, and OpenAI-hosted sandboxes at “standard container rates”. The Decoder’s write-up said billing “is based solely on token usage”, which is not what the documentation says. Web search, for example, costs $10 per 1,000 calls plus content tokens.
Hosted sandbox compute
Container rates are $0.03 for 1 GB, $0.12 for 4 GB, $0.48 for 16 GB and $1.92 for 64 GB per 20-minute session, billed by the minute with a five-minute minimum. The Agents API pages we read do not say which memory tier a hosted session uses.
A hosted sandbox can be deleted after an hour without activity or keep-alives, and that timeout “isn’t configurable”. The documentation says connected sandboxes receive keep-alives between turns, but it does not say whether that idle time is billed. Ask before running agents that wait on people.
Model choice moves the bill most
OpenAI’s observability guide uses an example model call of 5,000 input tokens, 1,500 of them cached, and 900 output tokens. Priced at the list rates for short context, the same call varies 45-fold by model. An agent task makes many such calls, and subagents make their own.
| Model | Input, cached, output per million | One example call | 1,000 such calls |
|---|---|---|---|
| GPT-6 Astra | $10, $1, $50 | $0.0815 | $81.50 |
| GPT-5.6 Sol | $4, $0.40, $20 | $0.0326 | $32.60 |
| GPT-5.6 Terra | $2, $0.20, $12 | $0.0181 | $18.10 |
| GPT-5.6 Luna | $0.20, $0.02, $1.20 | $0.0018 | $1.81 |
Two caveats apply. Sol’s price is promotional “at least through November 21, 2026”. And Astra charges $12.50 per million tokens for cache writes, which the Agents API usage fields “do not expose”, so its real cost per call can be higher.
Documented Agents API limits and defaults
These are the figures in the beta documentation on 11 September. They may change before general availability.
| Setting | Documented value |
|---|---|
| Hosted network access | Enabled by default; restricted mode allows 1 to 100 exact host names |
| Subagents running at once | 6 by default, not counting the coordinator |
| Function tools in subagents | Not supported |
| Files when creating a session | 50 per request; inline uploads 5 MiB each and 10 MiB in total |
| Published artifacts | 200 MiB per file, 500 MiB published together |
| Self-hosted connection wait | Up to five minutes, then the submission fails |
| Trace retrieval and exporters | Not part of the public beta API |
| Data residency | United States only |
| Zero Data Retention | Not supported, including with a self-hosted sandbox |
Defaults that deserve a second look
Two defaults lean permissive, which makes them a cybersecurity decision as much as a configuration one. Outbound network access is on unless you switch it off, although OpenAI’s own report example sets it to “disabled”. Subagents default to six at once, while the launch example sets three and the overview sets four. Subagents share the coordinator’s files, because “creating a subagent does not create another environment”. For self-hosted compute, “deleting a session neither stops its environment nor emits a deletion webhook”, so orphaned machines are your problem.
Usage figures are “not a final bill”
Session and turn objects report usage on a best-effort basis. The guide warns that usage “can be null when unknown”, that “missing usage does not mean zero usage”, and that “these counts are not a final bill”. Spending controls therefore belong in your billing dashboard and project limits, not in code that trusts the session object.
GPT-Live-1 and the Agents API Do Not Share Data Controls
OpenAI presents GPT-Live-1 and the Agents API as complementary, and technically they are. The data-controls table in OpenAI’s platform documentation shows they sit in different retention categories.
The GPT-Live-1 retention table
This comparison uses the endpoint rows from OpenAI’s “Your data” guide, plus the residency statements on each product’s own documentation.
| Control | GPT-Live-1 sessions | Agents API | Realtime API |
|---|---|---|---|
| Used for training | No | No | No |
| Abuse monitoring retention | 30 days | 30 days | 30 days |
| Application state retention | None, or 30 days if stored | Until deleted | None |
| Zero Data Retention eligible | Yes, with limitations | No | Yes |
| Processing regions | United States, Europe (EEA and Switzerland) | United States only | United States, Europe (EEA and Switzerland) |
Zero Data Retention stops at the delegation boundary
OpenAI’s GPT-Live-1 data note says it plainly: “Delegated backend models and tools have their own data controls.” An organisation with Zero Data Retention enabled can run the voice layer without stored application state. If that voice layer hands work to an Agents API session, the agent’s state is kept until deleted. The guide warns that ineligible endpoints “may retain application state when used, even if you have Zero Data Retention enabled”, and the Agents API overview adds that a self-hosted sandbox “does not make the Agents API ZDR-eligible”.
A European voice, an American agent
GPT-Live-1 can be processed in Europe. The Agents API “currently supports data residency only in the United States”. A European deployment can therefore keep the conversation in-region but not an Agents API backend. The UK sits outside that European processing region: OpenAI lists the UK for regional storage only. Regional processing also carries a 10% price uplift for eligible models released on or after 5 March 2026, so check each backend model separately.
Stored GPT-Live-1 recordings cannot be deleted by API
Session storage is off by default. When a project enables it, store: true keeps the completed recording for 30 days so it can be downloaded or forked into a new session. OpenAI’s data guide adds: “The API does not provide a public stored-session deletion endpoint.” With Zero Data Retention enabled, the setting is ignored and treated as false. Anyone recording customer calls needs to decide whether that 30-day window fits their privacy notice before switching storage on.
What GPT-Live-1 and Agents API Customers Claim
Both launch posts lean on customer quotes. We counted 13 and sorted them by whether they contain a measured outcome. Every figure is self-reported in OpenAI’s posts, with no published method behind it.
13 quotes, five measured outcomes
| Company | Product | Claim | Measured outcome |
|---|---|---|---|
| EliseAI | GPT-Live-1 | Code base 80% smaller, 23,000 lines removed | Yes |
| Speak | GPT-Live-1 | Almost 80% fewer interruptions during thinking pauses | Yes |
| Yelp | GPT-Live-1 | “Meaningful improvements in call handling rates” | No figure |
| Fin | GPT-Live-1 | Voice support moving toward “the natural flow of a phone call” | No |
| Cognition | GPT-Live-1 | Working with Devin by voice starts to feel like collaborating with a teammate | No |
| Ciridae | Agents API | Evaluation score 0.71 to 0.85, four times lower latency | Yes |
| SafetyKit | Agents API | 60% lower cost per case | Yes |
| Hypha | Agents API | 86% fewer failed agent responses | Yes |
| Long Lake | Agents API | Agents stood up “in hours” | No |
| WithCoverage | Agents API | Replaced prompt chains and custom tool handling | No |
| Dwelly | Agents API | Fanned work out across hundreds of agents | Scale, not outcome |
| deepsense.ai | Agents API | “Very strong” engineering quality in a live repository | No |
| Nash | Agents API | Thousands of long-running logistics agents | Scale, not outcome |
The Agents API quotes are more quantified than the GPT-Live-1 ones: three of eight against two of five. Hypha’s 86% is the most interesting number for architects, because it credits “separating the agent harness from the sandbox” rather than a better model.
Yelp’s million calls, without an improvement figure
Yelp says its restaurant voice agent, Yelp Host, has handled more than 1 million calls since launching in October 2025, and that GPT-Live-1 now runs the voice layer of Yelp Host and Hatch, the lead-management platform Yelp agreed to buy in January. Its release claims better call handling and fewer call transfers in production testing. As PPC Land noted, it attaches no figures to either. The underlying business logic did not change; only the voice layer was swapped.
OpenAI’s own phone line
The most concrete production number sits in an older post. OpenAI Presence, launched on 22 July for enterprise voice and chat agents, runs OpenAI’s own English-language support line. OpenAI says it “resolves 75% of inbound issues without human assistance” and that its improvement loop “reduced human handoffs by 15 percentage points in just 10 days”. The GPT-Live-1 announcement names Presence as another way to build on the model, sold through account directors rather than the public API.
Migrating a Realtime Voice Agent to GPT-Live-1
Existing Realtime API customers are the obvious first movers, and OpenAI’s migration guide is candid that this is a rebuild of the voice loop, not a model swap.
A new endpoint and new event names
GPT-Live-1 works only on v1/live/sessions, so Realtime session code has to change. Audio sent with input_audio_buffer.append becomes session.input_audio.append, and output arrives as session.output_audio.delta. Manual audio commits and turn triggers go, because “GPT-Live decides when to speak”. There is also no event marking the end of each spoken response, so speaking indicators must follow playback on the client.
Split one prompt into two
The guide tells teams not to copy a Realtime prompt wholesale into GPT-Live-1. Conversation style and delegation guidance go in the voice prompt; business rules and tool instructions go to the backend. It also reminds developers that “prompt instructions guide the models; they do not enforce those checks”. Confirmation and permission checks stay in application code, exactly where a careful AI agents deployment already keeps them.
Voicemail detection needs a second session
Outbound calling has a catch. In client mode, the delegation event carries metadata and timing “without raw audio”, and “a delegated backend does not automatically receive the waveform”. To detect an answering machine, OpenAI suggests running a separate Realtime session alongside GPT-Live-1 for part of the call. That means a second, token-billed meter on every outbound call until detection ends.
GPT-Live-1 mini did not make the API
In July OpenAI rolled out two versions to ChatGPT, GPT-Live-1 and GPT-Live-1 mini, and said it planned to bring “them” to the API. The system card says the mini model is the default for free ChatGPT users. On 10 September the price list named only gpt-live-1, and OpenAI’s model page for the mini version returned a 404. There is no cheaper GPT-Live tier for developers yet.
Where This Leaves DevDay and Agent Builder
The Agents API landed in a busy calendar. It changes what OpenAI has left to announce on 29 September, and it sharpens a deadline for teams still on Agent Builder.
Managed agents arrived 19 days early, under another name
On 7 September a TestingCatalog report said OpenAI was preparing “Managed Agents” for DevDay. Three days later the Agents API shipped, 19 days before the keynote. When we looked at that leak on 8 September, the leak never mentioned Codex. The product that shipped describes itself in its first line as built “with the Codex harness”.
Agent Builder’s 81-day countdown
OpenAI’s AgentKit page carries a 3 June update: Agent Builder and Evals “will no longer be available” from 30 November 2026. The recommended successors were the Agents SDK for code and Workspace Agents for prompting. The Agents API is now a third route out.
| Date | Event |
|---|---|
| 15 April 2026 | Agents SDK update adds sandbox execution with seven providers |
| 3 June 2026 | OpenAI announces Agent Builder and Evals will close on 30 November |
| 8 July 2026 | GPT-Live-1 and GPT-Live-1 mini launch in ChatGPT, backed by GPT-5.5 |
| 22 July 2026 | OpenAI Presence launches for enterprise voice and chat agents |
| 31 July 2026 | SynthID watermarking added to supported GPT-Live audio |
| 4 August 2026 | GPT-Live system card safety results corrected and re-run |
| 10 September 2026 | GPT-Live-1 reaches the API; Agents API enters public beta |
| 29 September 2026 | OpenAI DevDay at Fort Mason, San Francisco |
| 30 November 2026 | Agent Builder and Evals no longer available |
What to watch for on 29 September
Five open questions would change the picture. First, a general-availability date for the Agents API. Second, Zero Data Retention or European processing for agent sessions. Third, GPT-Live-1 mini in the API. Fourth, disclosure of the container tier behind hosted sandboxes. Fifth, trace export for agents. OpenAI said it will “iterate quickly” during the beta, so the documentation is worth re-reading after the keynote.
Should Your Business Build on GPT-Live-1 and the Agents API?
Neither product is a toy, and both are early. The right answer depends less on the models than on your data rules and how much of the loop you want to own.
Where GPT-Live-1 fits now
Inbound phone lines with bounded tasks are the clearest fit: bookings, order status, appointment changes and first-line triage. Those are the calls where interruptions and background noise break a turn-based agent. Keep the backend simple, keep confirmations in code, and design the handoff to a person before you design the greeting. The banking result above is a reminder of why.
Where the Agents API fits now
The Agents API suits internal, long-running work on data that may be processed in the United States: code review, research, report generation and incident triage. It is also a reasonable landing place for teams leaving Agent Builder before 30 November. A strong vendor management process should treat it as a beta service, with pricing and limits that can move.
When to wait for general availability
Hold back if you need Zero Data Retention across the whole voice-and-agent chain, if the agent layer must process data in Europe, or if your auditors expect traces exported to your own tools. Those are documented gaps today, not guesses.
Five checks before a GPT-Live-1 pilot
- Map every piece of customer data the voice layer will pass to a backend, and check that backend’s own retention row.
- Price a realistic call, including silence and hold time, against both GPT-Live-1 and a Realtime model.
- Confirm your usage tier’s concurrent-session cap covers your busiest hour.
- Decide whether recordings will be stored, knowing there is no API endpoint to delete them.
- Write the interruption rules down: which backend actions can be cancelled, and which must be reconciled.
GPT-Live-1 and Agents API: Frequently Asked Questions
What is GPT-Live-1?
GPT-Live-1 is OpenAI’s full-duplex voice model. It listens and speaks at the same time, handles interruptions and background noise, and hands reasoning and tool use to a separate backend model or agent. It launched in ChatGPT on 8 July 2026 and in the API on 10 September 2026.
How much does GPT-Live-1 cost?
GPT-Live-1 costs $0.05 per minute of session time, billed per second, which is $3.00 per hour. Silence and backend waiting time count. Backend model and tool usage is billed separately at normal API rates.
Is the Agents API free to use?
There is no platform fee, but it is not free. You pay the chosen model’s token rates, standard rates for built-in tools such as web search, and container rates for OpenAI-hosted sandboxes. Self-hosted sandboxes run on your own compute bill.
Can GPT-Live-1 run with Zero Data Retention?
Yes. GPT-Live-1 sessions are eligible for Zero Data Retention, and storage is then treated as off. The Agents API is not eligible, so a voice agent that delegates to it does not keep Zero Data Retention end to end.
Is GPT-Live-1 mini available in the API?
Not as of 11 September 2026. OpenAI’s price list names only gpt-live-1, and the model page for the mini version returned a 404, even though OpenAI said in July it planned to bring both models to developers.
What happens to Agent Builder?
OpenAI says Agent Builder and Evals will no longer be available from 30 November 2026. It recommends the Agents SDK for code-based workflows and Workspace Agents for prompt-based ones. The Agents API is now a third option for hosted, long-running agents.
References and Further Reading
Build more natural voice experiences with GPT-Live-1 in the API (OpenAI)
Introducing the Agents API (OpenAI)
GPT-Live System Card (OpenAI Deployment Safety Hub)
GPT-Live 1 model page (OpenAI API documentation)
Getting started with GPT-Live (OpenAI API documentation)
Voice agent cost optimisation (OpenAI API documentation)
Migrate to GPT-Live (OpenAI API documentation)
Agents API overview (OpenAI API documentation)
OpenAI-hosted sandboxes (OpenAI API documentation)
Agents API multi-agent guide (OpenAI API documentation)
Agents API observability and usage (OpenAI API documentation)
Data controls in the OpenAI platform (OpenAI API documentation)
Codex open-source repository (GitHub)
The next evolution of the Agents SDK (OpenAI)
Introducing OpenAI Presence (OpenAI)
Yelp and Hatch put OpenAI’s GPT-Live-1 into voice AI after 1 million calls (PPC Land)
OpenAI’s new Agents API gives developers the infrastructure behind Codex and ChatGPT (The Decoder)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.